Skip to content

Correlation and Diversification: The Actual Math

Twenty stocks that move together carry almost as much risk as one. The covariance term decides the rest - here is the two-asset formula worked line by line, plus a table across six correlations.

By the editorial team at The Long View How we check our guides 9 min read
Three printed line charts fanned out on a desk, one showing two lines moving in opposite directions
Three printed line charts fanned out on a desk, one showing two lines moving in opposite directions

Diversification is usually taught as a rule of quantity: hold more things, and risk falls. That rule is incomplete in a way that matters. Twenty stocks whose prices move together carry almost as much risk as one of them, and two assets that move in opposite directions can do more for a portfolio than twenty that move in lockstep. The variable that decides which happens is correlation, and it appears exactly once in the portfolio risk formula - in the term most people skip.

This article works the numbers end to end. Every figure below is arithmetic you can reproduce in a spreadsheet from the assumptions stated, which is the point: correlation is not a metaphor, it is a coefficient between -1 and +1 that you multiply into a covariance term.

The Two-Asset Formula, Term by Term

For two assets with weights w₁ and w₂, standard deviations σ₁ and σ₂, and correlation ρ, the portfolio variance is:

σp² = w₁²σ₁² + w₂²σ₂² + 2·w₁·w₂·ρ·σ₁·σ₂

The first two terms are each asset's own variance, scaled by its weight squared. The third term is the covariance: it can be positive, zero, or negative, and it is the only term that can reduce total risk without reducing expected return. When ρ = 1 the covariance term is at its maximum and the portfolio's standard deviation equals the weighted average of the two assets. When ρ is below 1, the portfolio's standard deviation is lower than that weighted average, and the gap is the diversification benefit.

Worked Example: 60/40 With Real Numbers

Assumptions, stated up front: Asset A has an expected return of 8 percent and a standard deviation of 15 percent; Asset B has an expected return of 4 percent and a standard deviation of 6 percent; the weights are 60 percent A and 40 percent B; the correlation between them is 0.2. These are round numbers chosen for arithmetic, not a forecast.

StepCalculationResult
Weighted average σ0.6 × 15% + 0.4 × 6%11.40%
w₁²σ₁²0.36 × 0.02250.008100
w₂²σ₂²0.16 × 0.00360.000576
Covariance term2 × 0.6 × 0.4 × 0.2 × 0.15 × 0.060.000864
Portfolio variancesum of the three terms0.009540
Portfolio σ√0.0095409.77%
Expected return0.6 × 8% + 0.4 × 4%6.40%

The portfolio's standard deviation is 9.77 percent against a weighted average of 11.40 percent. Nothing about the assets changed; the 1.63 percentage point reduction is entirely the effect of correlation being 0.2 instead of 1.0. The expected return of 6.40 percent is unaffected - diversification does not dilute return in this model, it reduces the dispersion around it.

The Same Assets at Different Correlations

Hold the weights and the two assets fixed and vary only ρ. This table is the whole argument for caring about correlation:

CorrelationPortfolio σSharpe (rf 3%)Geometric return
+1.011.40%0.2985.75%
+0.510.41%0.3275.86%
+0.29.77%0.3485.92%
0.09.31%0.3655.97%
-0.28.84%0.3856.01%
-0.58.07%0.4216.07%

Two things are worth noticing. First, the relationship is not linear: the first drop from 1.0 to 0.5 buys 0.99 points of risk reduction, while the last drop from 0.0 to -0.5 buys only 0.77. Second, the geometric return column moves even though the arithmetic expected return is fixed at 6.40 percent, because volatility drag is subtracted as σ²/2. At ρ = +1 the drag is 0.65 points; at ρ = -0.5 it is 0.33 points. Compounded over twenty years, that difference in drag is worth more than the difference in fees most people argue about.

Why Twenty Holdings Is Not a Number

Apply the formula to an equal-weighted basket of twenty stocks, each with a 30 percent standard deviation. If the average pairwise correlation is 0.5, portfolio σ is 21.74 percent. If the correlation were 0, the same twenty stocks would produce 6.71 percent. The number of holdings did not change at all; only the correlation did.

This is why "how many funds do you own" is the wrong question, and "what do these funds actually hold" is the right one. Two large-cap index funds from different providers typically hold overlapping names and correlate above 0.95 with each other - adding the second fund adds cost and complexity, not diversification. A sector fund whose top holdings are already the largest positions in your broad index fund adds concentration while appearing to add breadth.

Extending to Three Assets

The formula generalises: every pair contributes a covariance term. Take weights of 50, 30, and 20 percent; standard deviations of 15, 6, and 10 percent; expected returns of 8, 4, and 6 percent; and correlations of 0.2 between A and B, 0.6 between A and C, and 0.1 between B and C. The weighted average standard deviation is 11.30 percent; the portfolio standard deviation is 9.36 percent. The expected return is again 6.40 percent, and the Sharpe ratio at a 3 percent risk-free rate is 0.363.

The third asset adds less than the second did, for the same reason that the twentieth stock adds almost nothing: what matters is the correlation of the marginal asset with the existing portfolio, not the count. An asset that correlates 0.9 with everything you already own is a small improvement even if it is a good asset in isolation.

Estimating the Number Yourself

The formula is only as good as the ρ you feed it, and a plausible number beats a precise-looking one. A workable method with free tools:

  1. Download monthly total-return data for both assets, five years is a reasonable compromise between recency and sample size. Sixty monthly observations give a correlation estimate with a standard error of roughly 0.1, which is worth knowing before you treat 0.2 and 0.4 as meaningfully different.
  2. In a spreadsheet, put the two return series in adjacent columns and use CORREL on the monthly returns. Monthly data captures more independent observations than annual data; daily data overweights noise from thin trading and time-zone effects.
  3. Compute the same correlation over two sub-periods, such as the first half and second half of the window. If the two estimates differ by more than about 0.3, the relationship is unstable and you should plan for the higher of the two.
  4. Convert annual standard deviations to the same basis as your returns before using them: σ_annual = σ_monthly × √12. Mixing monthly and annual figures is the most common arithmetic error in this calculation.

One trap deserves its own sentence: correlations computed on price levels rather than returns will be near 1 for almost any two growing series, which is an artefact of the trend, not a fact about the assets. Use returns, and use total returns that include distributions.

Where the Model Breaks

Three limitations are worth stating plainly, because they are the difference between using this framework and trusting it.

  • Correlations are estimated, not known. The number you put in the formula is a historical average with a wide error bar. Long-run stock-to-stock correlations in diversified markets are commonly estimated around 0.5 to 0.7; stock-to-bond correlations have historically sat near zero but turned positive during the 2022 inflation shock, when both fell together.
  • Correlations rise in a crisis. In a liquidity event, forced selling hits everything at once and pairwise correlations converge toward 1. The diversification that worked in calm markets is least available exactly when it is most needed - which is why the model understates tail risk rather than overstating it.
  • Standard deviation is not the same as risk for everyone. A drawdown you cannot tolerate is a different failure from a volatility number, and this model cannot see your time horizon or your need for the money.

None of these makes the arithmetic wrong; they make it a tool with a stated domain. The practical response is to prefer assets whose low correlation has an economic reason - different currencies, different growth drivers, different parts of the capital structure - rather than assets that merely had a low measured correlation over the last three years.

What This Means for a Real Portfolio

The formula converts a vague instruction into a checklist. When considering an addition, ask what its correlation with the existing portfolio is likely to be, and whether that correlation has a reason to persist. When a fund is added for "diversification", check the overlap in its top ten holdings against what you already own - an overlap above roughly half is a signal that you are buying the same exposure twice.

Rebalancing is the mechanism that harvests the benefit. A portfolio whose assets drift apart in performance quietly becomes concentrated in whatever went up, which raises both its risk and its correlation with itself. Returning to target weights at a fixed interval or at a band breach is the disciplined way to sell what has become expensive and buy what has become cheap. The mechanics are covered in the rebalancing bands guide, and the wider framework is in asset allocation.

Two related pieces complete the picture: why bonds behave the way they do when rates move, which explains the economic reason stock-bond correlation changes, and the Sharpe ratio walkthrough, which shows what to do with the risk number once you have it. For the volatility-drag arithmetic used above, the real return guide extends the same compounding logic.

This is educational material, not investment advice. It does not account for your circumstances, and it does not recommend any security or fund. The assumptions in the worked example are illustrative round numbers; substitute your own and the arithmetic still holds.